Boris demands TfL agree to Driverless Trains part 2

[alpinejohn]

Just how long will it be before someone drops a phone under/through that lash-up and decides to retrieve said precious item and ends up getting caught in the fishing net and garroted. In a crowded underground platform situation it is a recipe for disaster and lawsuits. Next...

OK you do not need to gold plate door systems - indeed there is no reason why door systems should not be made of far cheaper solid sheet material like ply or sheet steel. The big cost would however remain all those door opening mechanisms and related safety sensors. I would be astonished if bulk buying would allow you to cut up and running costs down by 10%. Presumably TFL have pretty recent bills for the stuff on the Elizabeth line so a few back of the envelope calculations should give you prices for installations at Waterloo and Bank.

Sadly virtually all stations are unique so virtually every station would need customisation to fit PEDs - so you could not get a massive discount if you immediately decided to fit PEDs to every LUL platform.

Fundamentally too many people seem to think an automated railway could easily be made as safe as an automated lift - but even lifts need maintence and reliable power supplies and they still break down often with the fire brigade turning up - sometimes hours later to release the victims. I would hate to see people emerging from a trapped train after spending several hours in the dark.

But it would be far from simple to automate the Waterloo and City and it would be massively costly - just to eliminate a handful of drivers wage packets with those savings largely offset by the need for extra train captains and extra CCTV monitoring staffing.

I fear this is another piece of total madness from someone who will never be held to account for his gross misspending of public monies.

Yes the Waterloo and City line is as simple as it can get, but even that has sidings and maintenance activity which cannot be automated. There are also a rabbit warren of supporting systems and tunnels which would all need someone to ensure are locked and monitored at least with CCTV 24/7 to stop lunatics up to mischief, and probably it would mean a lot more cleaning staff would be required to keep the place from turning into a dump.

Inherently if your real objective is to make the undergound system "unions/strike proof" that means eliminating ALL the staff from ALL activities. So no one on trains or platforms, no one doing maintenance and no one in control rooms. Just leaving it all up to a micro-chip which will probably stop working if the power ever fails.

A cardinal rule of risk assessment is to think outside the box and never underestimate the damage which can be wreaked by someone with a JCB. Hence this focus on automation regardless of costs or benefits, poses a risk which I, and I suspect many others users will prefer not to take.

[billbedford]

Aug 26, 2021 7:46:58 GMT Chris L said:

Aug 26, 2021 7:39:49 GMT brigham said:

" They're under starter's Orders...

...and they're off!"

Imagine that on the already narrow Victoria line platforms. (The lighting runs above the platform edge might get in the way.)

There's a range of different platform edge barriers in use on Japanese railways shown on this video

[class411]

Aug 26, 2021 8:24:27 GMT alpinejohn said:

Just leaving it all up to a micro-chip which will probably stop working if the power ever fails.

There is no probably about it.

Micro-chips are extremely militant. Stop paying their wages and they are instantly 'out'.

[d7666]

Aug 25, 2021 17:16:23 GMT class411 said:

Aug 25, 2021 15:40:14 GMT d7666 said:

A good start is to just wire along Oxford Street, and require that any bus that runs anywhere along it must be a trolley equipped battery bus, and it must run under the wires (except to overtake***).

The problem there, surely, is that you would need a vast number of buses (because that's a piece of road that forms a part of many routes), equipped with pick ups and batteries, that would also need to be prime movers, because the rest of their route is not electrified, and all battery buses are not yet practical.

You don't need any more buses than the total number of n=buses used now.

Policy already is to eliminate all hydro carbon fuel vehicles eventually. Which, at the moment, means batteries for all practical purposes. So you still need the same number of battery buses as you will need anyway.

All battery buses need charging somehow. If the default charge method is overhead contacts _ like a lot of operators already use _ you are not adding anything new or different.

If the central section [oxford street] of these routes is electrified, batteries will, get a charge while using it. This extends the useful mileage range.

The whole point IS to create a situation of needing to extended wires if battery operation won't suffice. You need a critical mass to catalyse the rest.

PS - remember my comment in another thread recently, as soon as some one suggests something innovative, everyone finds 1000 reason not to do it. QED.

[class411]

Aug 26, 2021 15:19:59 GMT d7666 said:

PS - remember my comment in another thread recently, as soon as some one suggests something innovative, everyone finds 1000 reason not to do it. QED.

We've been asked not to talk about buses, so all I'll say on that is that I can see your points.

On the bit quoted: Yes, that's exactly how it should be. If, for example, you propose driverless trains, everybody should be coming up with reasons why they won't work until all those points have been addressed.

It is just (as I well know from experience) damnably annoying if it's your own brilliant idea that everyone if picking to pieces.

[d7666]

Driverless trains, or varying Edgware Road District destinations other than Wimbledon, or whatever is being debated, perhaps I was not clear.

Everyone comes up with "but we've always done it that way" reasons against new ideas - no-one [almost no-one] comes up with supporting suggestions or varying alternative options.

[Chris M]

Given that all change costs and continuing with the status quo costs (relatively) nothing, it is up to the supporters of new ideas to show all of:

• There is a problem with status quo or is somehow suboptimal
• The new idea solves that problem or is in some other way objectively better
• The new idea either:
  
• doesn't introduce new problems
• introduces only problems that can be fixed or worked around trivially
• introduces only problems that are significantly minor compared to both the current situation and the benefits of the new idea

[/ol][li]The costs of introducing the new idea are reasonable[/li][li]The benefits of introducing the new idea significantly outweigh the costs of doing so[/li][/ul]In the case of driverless trains, on the Underground it has been established in detailed study after detailed study that the costs are enormous and the benefits tiny. In the case of Wimbledon to Edgware Road it's not clear that any of the ideas presented will both solve the problem and not introduce equal or worse problems elsewhere (I'm not even certain that all the ideas will even solve the problem).

It is entirely reasonable for people to ask supporters of new ideas to answer questions about their ideas and to demonstrate that they do indeed do all of the above. If someone comes up with a potential problem with the new idea the correct responses are one of "That's not actually a problem because...", "That problem is solved by...", "That is a problem, but it can be mitigated by...", "I'm not sure about that, let me do more research" or "Ah yeah, you're right, that is a blocker.". Note "Why are you always opposed to new ideas?" and "Well you come up with a new idea them!" are not in that list.

[class411]

Aug 28, 2021 15:09:35 GMT d7666 said:

Driverless trains, or varying Edgware Road District destinations other than Wimbledon, or whatever is being debated, perhaps I was not clear.

Everyone comes up with "but we've always done it that way" reasons against new ideas - no-one [almost no-one] comes up with supporting suggestions or varying alternative options.

Don't forget that not everyone here knows enough about LU operations to make sensible suggestions.

Given that, here are a couple I've thought about over the years:

1) Train dispatch

Have video cameras pointing at every doorway/danger point on the network. Use AI to determine if the doorways is definitely clear - if not alert a human. A central control room would house enough people to dispatch as many trains as were likely to want to depart at one time.

2) Train operation

Use video cameras to monitor multiple parts of the train and its environment. If the automatic system fails for any reason, drivers situated at a central control point could either rectify the problem, or, failing that, crawl the train forward to somewhere it could be sorted out.



Both options that would throw up all sorts of obstacles and objections, I have no doubt.

[jimbo]

Aug 28, 2021 17:45:29 GMT class411 said:

Don't forget that not everyone here knows enough about LU operations to make sensible suggestions.

Given that, here are a couple I've thought about over the years:

1) Train dispatch

Have video cameras pointing at every doorway/danger point on the network. Use AI to determine if the doorways is definitely clear - if not alert a human. A central control room would house enough people to dispatch as many trains as were likely to want to depart at one time.

2) Train operation

Use video cameras to monitor multiple parts of the train and its environment. If the automatic system fails for any reason, drivers situated at a central control point could either rectify the problem, or, failing that, crawl the train forward to somewhere it could be sorted out.

Both options that would throw up all sorts of obstacles and objections, I have no doubt.

I think these both sound like part of TfL planning for possible driverless operations with coming train deliveries. Currently the driver does these things with the help of CCTV. 

[billbedford]

Aug 28, 2021 17:45:29 GMT class411 said:

Don't forget that not everyone here knows enough about LU operations to make sensible suggestions.

Given that, here are a couple I've thought about over the years:

1) Train dispatch

Have video cameras pointing at every doorway/danger point on the network. Use AI to determine if the doorways is definitely clear - if not alert a human. A central control room would house enough people to dispatch as many trains as were likely to want to depart at one time.

2) Train operation

Use video cameras to monitor multiple parts of the train and its environment. If the automatic system fails for any reason, drivers situated at a central control point could either rectify the problem, or, failing that, crawl the train forward to somewhere it could be sorted out.



Both options that would throw up all sorts of obstacles and objections, I have no doubt.

Is it really a good idea to solve social/political problems with technology?

[class411]

Aug 28, 2021 21:01:27 GMT jimbo said:

I think these both sound like part of TfL planning for possible driverless operations with coming train deliveries. Currently the driver does these things with the help of CCTV. 

I know that there is a lot of CCTV, now, but I have never seen it mooted that it could be used for remote train dispatch or operation.

Aug 29, 2021 8:06:36 GMT billbedford said:

Is it really a good idea to solve social/political problems with technology?

That depends on the circumstances. In this case, my personal opinion would be: no. But that doesn't mean discussing the technicalities is not interesting.

Also, my first suggestion could improve safety, as every door would be clearly visible, and the AI would be able to detect if anything outside the train (and close) to it was accelerating at the same speed as said train and, if it was halt the departure.

[Chris M]

Aug 29, 2021 8:45:54 GMT class411 said:

Also, my first suggestion could improve safety, as every door would be clearly visible, and the AI would be able to detect if anything outside the train (and close) to it was accelerating at the same speed as said train and, if it was halt the departure.

I'm not sure how well CCTV focused on each door would do for trap and drag incidents - for detection of clear doorways, etc you would want cameras at approximately 90° to the train, looking downwards from above the heads of passengers on the platform. Anything trapped would rapidly go out of view of each camera, and while it would be possible (at least in theory) to link them up somehow that would make detection a lot harder computationally. You would want something looking lengthways along the train, so you would need a different set of cameras. However I think improving door sensitive edge technology is going to prove far better value for trap and drag type incidents.

Any AI used for anything safety critical is going to need to be trained extremely rigorously, first on recorded training data and then observing real world data and offering its opinions for a human to observe, then it will need to do the job with human supervision before being let loose. It might be the case that each platform needs training individually or in small groups of similar platforms. Certainly PEDs and non-PEDs will need training differently, as will likely curved and straight platforms. All this is going to take a huge amount of both time and money, so the question needs to be asked whether it is worth it? I don't know the answer to that.

[spsmiler]

Aug 28, 2021 15:09:35 GMT d7666 said:

Driverless trains, or varying Edgware Road District destinations other than Wimbledon, or whatever is being debated, perhaps I was not clear.

Everyone comes up with "but we've always done it that way" reasons against new ideas - no-one [almost no-one] comes up with supporting suggestions or varying alternative options.

Well yes, if Earls Court could be unclogged then both High Street Kensington and Edgware Road could be used for more services to more destinations.

In addition, if CBTC increases capacity on the southern side of the Circle then maybe more trains could be run between Edgware Road and (for instance*) Plaistow, boosting the Circle line service between Victoria and Paddington. This would not directly affect Earls Court! But capacity issues as Edgware Road might mean that some Wimbledon trains would need to stop at High Street Kensington.

*Had it been available Mansion House could have been a better terminus point as the section beyond Aldgate East might not be able to accept more trains.

re: train AI and despatch, yes this would likely 'see' people caught in doors, probably even satchels, rucksacks / backpacks, arms, etc (especially as 'doors properly closed' detection would also flag an issue in these situations) but what about long hair, scarves, belts and winter coat cords / toggles? As a young child my gloves were on elastic so that if I took them off I did not loose them... these hung down from my jacket sleeves and could also be easily caught in closed doors.

[class411]

Aug 29, 2021 10:24:49 GMT Chris M said:

I'm not sure how well CCTV focused on each door would do for trap and drag incidents - for detection of clear doorways, etc you would want cameras at approximately 90° to the train, looking downwards from above the heads of passengers on the platform. Anything trapped would rapidly go out of view of each camera, and while it would be possible (at least in theory) to link them up somehow that would make detection a lot harder computationally.

The cameras would be above the train pointing straight down, with wide angle lenses so that the whole length of the train was visible. I wouldn't say that the AI would be a trivial computational load, but, with the power of modern computers it would not be close to unimplementable.

but You would want something looking lengthways along the train, so you would need a different set of cameras.

Wouldn't work. Too many curved platforms.

Aug 29, 2021 12:04:20 GMT spsmiler said:

re: train AI and despatch, yes this would likely 'see' people caught in doors, probably even satchels, rucksacks / backpacks, arms, etc (especially as 'doors properly closed' detection would also flag an issue in these situations) but what about long hair, scarves, belts and winter coat cords / toggles? As a young child my gloves were on elastic so that if I took them off I did not loose them... these hung down from my jacket sleeves and could also be easily caught in closed doors.

My suggestion circumvented those problems because it detected things accelerating at the same rate as the train (in addition to clear space detection).

[Tom]

Aug 29, 2021 12:04:20 GMT spsmiler said:

re: train AI and despatch, yes this would likely 'see' people caught in doors, probably even satchels, rucksacks / backpacks, arms, etc (especially as 'doors properly closed' detection would also flag an issue in these situations) but what about long hair, scarves, belts and winter coat cords / toggles? As a young child my gloves were on elastic so that if I took them off I did not loose them... these hung down from my jacket sleeves and could also be easily caught in closed doors.

This is exactly how a young child was dragged to their death on one of the Piccadilly line platforms at Holborn in 1997. They were wearing a coat with a toggle, which got caught in the doors and couldn't be released.

[Chris M]

There is a big difference between an item of clothing trapped in the doors when the passenger is on the train and when the passenger is on the platform.
In the former case it's almost never a big deal - the passenger can release themselves the next time those doors open and/or use the emergency alarm.
In the latter case it is absolutely a big deal and the train needs to stop ASAP. Something trapped in the doors will only partially be accelerating at the same speed as the train, and that part may be tiny; everything else will be accelerating at a slower rate as the passenger will have more resistance from the platform, other passengers, etc. Initially any visual system will be almost completely unable to detect the difference between a passenger being dragged and a passenger walking quickly in the same direction as the train and something trapped from within the train. When it does detect the issue it needs to send a message to the train, which adds another delay. A sensitive edge system on the train will have detected a trap and drag almost immediately and stopped the train with no communication delay.

[class411]

Aug 29, 2021 20:10:04 GMT Chris M said:

In the latter case it is absolutely a big deal and the train needs to stop ASAP. Something trapped in the doors will only partially be accelerating at the same speed as the train, and that part may be tiny; everything else will be accelerating at a slower rate as the passenger will have more resistance from the platform, other passengers, etc.

Chris, if someone is effectively attached to the train, they can only accelerate at a different rate for a very small amount of time (unless they are attached by a bungee). If their d2v/dx² is different to the train they cannot be attached to it.

Initially any visual system will be almost completely unable to detect the difference between a passenger being dragged and a passenger walking quickly in the same direction as the train and something trapped from within the train.

Actually, not true. The motion signature of a person walking is completely different to that of something smoothly accelerating.

When it does detect the issue it needs to send a message to the train, which adds another delay.

I really don't think that a couple of milliseconds will make much difference.

A sensitive edge system on the train will have detected a trap and drag almost immediately and stopped the train with no communication delay.

I'm not sure what that has do do with anything. I wasn't suggesting that the system replaced the sensitive edge - merely that if you had a remote dispatch system a useful corollary benefit would be an extra level of monitoring of drag incidents.

[Chris M]

You are assuming that the person is trapped by something with absolutely no elasticity, that cannot partially unwind or tear, that does not have loose parts that can flap, that a person being dragged will not try and get themselves free and/or try keep up with the train (at least at first), that the passengers progress along the platform will not be impeded and that no person walking alongside a train, or anything they are carrying or wearing will never make movements that are similar to those of a person being dragged. I don't think any single one of these is a safe assumption, let alone two or more combined.

When you have a sensitive edge system that works reliably with existing technology with no communication or processing delays I just do not understand why you would spend the vast amounts of time and money required on an AI system that can only ever hope to equal it at best?

[class411]

Aug 29, 2021 23:47:10 GMT Chris M said:

You are assuming that the person is trapped by something with absolutely no elasticity, that cannot partially unwind or tear, that does not have loose parts that can flap, that a person being dragged will not try and get themselves free and/or try keep up with the train (at least at first), that the passengers progress along the platform will not be impeded and that no person walking alongside a train, or anything they are carrying or wearing will never make movements that are similar to those of a person being dragged. I don't think any single one of these is a safe assumption, let alone two or more combined.

None of those things are safe assumptions, but the system would use multiple approaches. I'm not going to go through your objections one by one, but, just as an illustration, I'll deal with two:

"assuming that the person is trapped by something with absolutely no elasticity"

If the person is attached with no elasticity, they will accelerate at almost exactly the same rate as the train.

If there are attached by something elastic (or with some slack), at some point the elasticity of slack will be exhausted. At that point their d³x/dt^{3 (rate of change of acceleration)}  will jump in a way that would not happen naturally.

Using these two facts, together with the constraint that the action must be taking place extremely close to the train, and you have an answer to one problem.


"passengers progress along the platform will not be impeded"

If the passengers progress is significantly impeded, the system will see three things  (assuming the collision does not break the attachment): Firstly, it will see an object, very close to the train accelerating, it will see two objects very close to the train collide, then it will see the impeding object (again very close to the train) accelerate unnaturally rapidly. That would be enough to halt the train.

It took me less that half an hour to come up with those two answers. IF LU decided to implement such a system, the would probably have a team working on it for a couple of months, who would probably come up with a set of much better answers.

When you have a sensitive edge system that works reliably with existing technology with no communication or processing delays I just do not understand why you would spend the vast amounts of time and money required on an AI system that can only ever hope to equal it at best?

You still seem to think that I'm suggesting that the camera system would be a replacement for sensitive edges.

I am not.

Everything fails from time to time, and door mechanisms are not exempt (nor are they perfect). All I was suggesting was that if you implemented a system for remote dispatch, you could leverage that to add an extra layer of safety which, because it worked in a completely different way to sensitive edge would not suffer from common mode failure.

[Chris M]

It is true that everything fails from time-to-time, but as I explained in the first response it seems very unlikely to me that cameras used for dispatch (requiring a close focus on doorways) will be suitable for this (which would require an overview of the platform-train interface along the length of the train). Even if the cameras are suitable for the task and you've managed to develop a viable remote dispatch system you would still have to spend the same amount of time and very nearly the same amount of money (basically everything but the cameras and system-to-train communication) training an AI to do this (even if you are using AI for dispatch, this would be a separate task with limited overlap) and I remain to be convinced that it could even plausibly be a worthwhile use of resources.

[class411]

Aug 30, 2021 10:48:03 GMT Chris M said:

It is true that everything fails from time-to-time, but as I explained in the first response it seems very unlikely to me that cameras used for dispatch (requiring a close focus on doorways) will be suitable for this (which would require an overview of the platform-train interface along the length of the train).

Cameras are so cheap, now, that even having one set for each function would not be bank breaking.

Even if the cameras are suitable for the task and you've managed to develop a viable remote dispatch system you would still have to spend the same amount of time and very nearly the same amount of money (basically everything but the cameras and system-to-train communication) training an AI to do this.

You would not 'train' the AI for a task like this.

AI is trained when it is implemented as a neural net. Both of these applications are completely unsuitable for neural net implementation. It would be programmed with specific logic.

(even if you are using AI for dispatch, this would be a separate task with limited overlap) and I remain to be convinced that it could even plausibly be a worthwhile use of resources.

Well, it's not you or I who have to be convinced.

I may not have convinced you that it's feasible, but I have spent decades building computer systems, as programmer, analyst, and systems architect, and you have not convinced me that it is not feasible - although it's not impossible that, if looked into more deeply, there might be something that makes it impractical to implement.

Anyway, we've nicely proven d766's contention that any idea will find someone willing to find many reasons not to do it. (Which is all good and part of the design process, IMO.)

[philthetube]

One cannot say that sensitive edge is 100% failsafe, I am not aware of any issues with it, and it is a very clever, and simple system. (clever as in well thought out, not intelligent).

The issue I see with using cameras in this way would be having them recognise when people were sticking out of trains because the train is full, or just standing on the step saying goodbye. There is currently discretion as to when the doors are closed, cameras would struggle to see if the doors were obstructed for a good reason or not.

[class411]

Aug 30, 2021 14:09:45 GMT philthetube said:

One cannot say that sensitive edge is 100% failsafe, I am not aware of any issues with it, and it is a very clever, and simple system. (clever as in well thought out, not intelligent).

The issue I see with using cameras in this way would be having them recognise when people were sticking out of trains because the train is full, or just standing on the step saying goodbye. There is currently discretion as to when the doors are closed, cameras would struggle to see if the doors were obstructed for a good reason or not.

That is why I stated, in my initial post on this, that the camera system would pass off anything other than a clear doorway to a human operator.

They would really be a system to reduce the workload of the real people. This would be a safety benefit because it would mean that the operators would only see doors where there was a possible obstruction, rather than having to repeatedly look at over a dozen doorways, most of which were clear.

Also, the system could ensure that the operators continued to monitor any door areas that were not 100% clear (as could happen on a crowded platform) as the train departed.

[bomo]

Aug 28, 2021 17:45:29 GMT class411 said:

Aug 28, 2021 15:09:35 GMT d7666 said:

Driverless trains, or varying Edgware Road District destinations other than Wimbledon, or whatever is being debated, perhaps I was not clear.

Everyone comes up with "but we've always done it that way" reasons against new ideas - no-one [almost no-one] comes up with supporting suggestions or varying alternative options.

Don't forget that not everyone here knows enough about LU operations to make sensible suggestions.

Given that, here are a couple I've thought about over the years:

1) Train dispatch

Have video cameras pointing at every doorway/danger point on the network. Use AI to determine if the doorways is definitely clear - if not alert a human. A central control room would house enough people to dispatch as many trains as were likely to want to depart at one time.

2) Train operation

Use video cameras to monitor multiple parts of the train and its environment. If the automatic system fails for any reason, drivers situated at a central control point could either rectify the problem, or, failing that, crawl the train forward to somewhere it could be sorted out.



Both options that would throw up all sorts of obstacles and objections, I have no doubt.

AI in its current stage of evolution, or indeed that likely in the foreseeable future, is simply incapable of assessing the nuance of risk to the same granularity as a human.  That's part of the reason why development of driverless road vehicles has stalled somewhat after all the hype about it a few years ago.  AI would of necessity have to take a very cautious approach to risk assessment which in a real world high volume system like transport means delays and lower throughput.  In my view the public are less likely to accept machines making mistakes resulting in injury or death than they are of humans making mistakes.

[class411]

Aug 31, 2021 8:17:19 GMT bomo said:

AI in its current stage of evolution, or indeed that likely in the foreseeable future, is simply incapable of assessing the nuance of risk to the same granularity as a human.  That's part of the reason why development of driverless road vehicles has stalled somewhat after all the hype about it a few years ago.  AI would of necessity have to take a very cautious approach to risk assessment which in a real world high volume system like transport means delays and lower throughput.  In my view the public are less likely to accept machines making mistakes resulting in injury or death than they are of humans making mistakes.

Again, I was suggesting this as an additional level of security, not a replacement.

Also, your contention that self drive cars are not yet reliable and thus a system such as I mooted is also not feasible is, I believe, flawed. A fully autonomous self driving car is an incredibly difficult proposition, requiring the system to take in vast amounts of data relating to many objects, both static and moving at multiple different speeds in multiple different directions. It also needs to be able to reliably determine the intentions of actors controlling some of these objects.

This system just needs to watch a thin strip and note if any object in it has an unusual acceleration profile that could be the result of something being dragged by the train.

There are already many systems that can automatically recognise objects, (think, for example, of face recognition and focus tracking that are already available in quite cheap cameras). You then need a layer on top of that to work out how they are moving. Certainly not trivial, but well within the abilities of a small, competent, team of programmers.

[castlebar]

l can foresee a situation that if we did have "driverless" ( = unstaffed) trains, so many people will be tasked to sit at desks looking at live videos of trains all around the system, one day, somebody will come up with the idea of putting staff on the trains to reduce staff numbers, and making all the video viewers, all sitting at desks, unnecessary. It's the law of unintended consequences. Think of the office space that would free up, or will these people be "working from home". (That is a potential disaster waiting to happen, e.g. "I'd just gone to the front door to collect a parcel when the accident happened").

It will only require an "incident" to take place whilst the monitor viewer has 'gone for a wee' or something or a monitor to fail at a critical time, (smoke for example), for there to (have to) be an expensive about turn and a re-think that it wasn't such a good idea after all.

[35b]

Aug 31, 2021 8:45:33 GMT class411 said:

Aug 31, 2021 8:17:19 GMT bomo said:

AI in its current stage of evolution, or indeed that likely in the foreseeable future, is simply incapable of assessing the nuance of risk to the same granularity as a human.  That's part of the reason why development of driverless road vehicles has stalled somewhat after all the hype about it a few years ago.  AI would of necessity have to take a very cautious approach to risk assessment which in a real world high volume system like transport means delays and lower throughput.  In my view the public are less likely to accept machines making mistakes resulting in injury or death than they are of humans making mistakes.

Again, I was suggesting this as an additional level of security, not a replacement.

Also, your contention that self drive cars are not yet reliable and thus a system such as I mooted is also not feasible is, I believe, flawed. A fully autonomous self driving car is an incredibly difficult proposition, requiring the system to take in vast amounts of data relating to many objects, both static and moving at multiple different speeds in multiple different directions. It also needs to be able to reliably determine the intentions of actors controlling some of these objects.

This system just needs to watch a thin strip and note if any object in it has an unusual acceleration profile that could be the result of something being dragged by the train.

There are already many systems that can automatically recognise objects, (think, for example, of face recognition and focus tracking that are already available in quite cheap cameras). You then need a layer on top of that to work out how they are moving. Certainly not trivial, but well within the abilities of a small, competent, team of programmers.

I have a problem with this concept, and it’s based on the failures of automatic cars that are “minded” by drivers. It takes a time to transfer attention fully to a situation; time that is less available when the person monitoring is not fully engaged in the task and situationally aware.

[class411]

Sept 1, 2021 8:17:37 GMT 35b said:

I have a problem with this concept, and it’s based on the failures of automatic cars that are “minded” by drivers.

I've always thought that that was an utterly idiotic idea (except for safety drivers, whilst testing). You remove the need for drivers to do anything for the vast majority of the time, and expect them to be instantly 100% involved the moment the automatic system makes an error.

The man is the Tesla that crashed into a lorry because the auto system thought it was the sky was, apparently, not paying attention.

Even if he had been, the scenario could well have been:

Man: Oh, look, a lorry.
Man: That's funny, shouldn't we be slowing down.
Man: Jeez, I'd better brake mysel...
Two large chunks of metal: ***BANG***

It takes a time to transfer attention fully to a situation; time that is less available when the person monitoring is not fully engaged in the task and situationally aware.

I'm not sure which concept you have a problem with.

The dispatch aid system, where it determines which doors are definitely clear (actually a trivial computational task) would increase safety, because it would mean that the operator would only be presented with situations that were potentially hazardous, rather than spending the majority of their time looking at situations that were not.

And the automatic drag detector would not involve humans.

[class411]

Aug 31, 2021 9:39:09 GMT castlebar said:

l can foresee a situation that if we did have "driverless" ( = unstaffed) trains, so many people will be tasked to sit at desks looking at live videos of trains all around the system, one day, somebody will come up with the idea of putting staff on the trains to reduce staff numbers, and making all the video viewers, all sitting at desks, unnecessary. It's the law of unintended consequences. Think of the office space that would free up, or will these people be "working from home". (That is a potential disaster waiting to happen, e.g. "I'd just gone to the front door to collect a parcel when the accident happened").

It will only require an "incident" to take place whilst the monitor viewer has 'gone for a wee' or something or a monitor to fail at a critical time, (smoke for example), for there to (have to) be an expensive about turn and a re-think that it wasn't such a good idea after all.

I love the idea that, some years after an automated, or remote, system is introduced, someone comes up with the idea that it would be cheaper or safer to actually have a person in the middle of the train to determine when it's safe for the train to depart and guard against drag and kill incidents.

BUT

If they are going to start to move towards driverless trains, they are going to have to implement some system for dispatching trains.

That will be one of:

a) Someone on every platform.
b) Someone monitoring (in some sense) remotely.
c) Some fully automatic system.

and they will need to have this in place before they even begin testing with passengers.

[Chris M]

Sept 1, 2021 8:44:24 GMT class411 said:

The dispatch aid system, where it determines which doors are definitely clear (actually a trivial computational task) would increase safety, because it would mean that the operator would only be presented with situations that were potentially hazardous, rather than spending the majority of their time looking at situations that were not.

A large part of the way that you know something is not right is because you have spent so long looking at situations where it is. I would very surprised if there is any experienced train operator who has never held off closing the doors because for some reason it just felt wrong before it became clear why. Obviously its impossible to prove that this did prevent an incident but it clearly will have done on at least some occasions.

If you are being pinged to attention by an AI, you first have to orient yourself, assess the scene and work out what is happening, even if this only takes 1-2 seconds that's time wasted. If the person being pinged is monitoring multiple locations the time taken will increase. Even if location information is present on screen for them, they still have to process it and switch to the new characteristics. People moving the opposite way along a platform to usual will be an obvious sign something is up to a t/op who intimately knows every platform and can't help but know where they are; but when someone is responsible for looking at multiple platforms they have to think about which one they're seeing and what is normal there.

If the AI is almost always correct, operators will get complacent and miss things when the AI isn't. If the AI is wrong too often, operators will learn to ignore its warnings (see "crying wolf").